UB55 multivariate analysis: hi-lo contrast, group-level

1 intro

Purpose:

1. Preliminary look at hi/lo multivariate contrast (cross-validated euclidean) in each task.
2. Preliminary qualitative comparison of effects of prewhitening and pattern standardization on multivariate contrast.

Notes on analyses:

Target TRs:

• Axcpt: 7, 8, 9
• Cuedts: 9, 10
• Stern: 11, 12
• Stroop: 2, 3, 4

Contrasts (cross-validated euclidean):

• Axcpt: \(\text{BY} - \text{BX}\)
• Cuedts: \((\text{InConSwitch} + \text{InConRepeat} - \text{ConSwitch} - \text{ConRepeat})/2\)
• Stern: \(\text{LL5RN} - \text{LL5NN}\)
• Stroop: \((\text{PC50InCon} + \text{biasInCon} - \text{PC50Con} - \text{biasCon})/2\)

Normalizations:

• “stand”: standardized cross-validated euclidean (comparable to linear correlation)
  • the across-vertex mean and standard deviation are removed from patterns prior to cross-validation
• “unsta”: unstandardized cross-validated euclidean (includes mean and scale information)
• “white”: spatially pre-whitened cross-validated euclidean (‘mahalanobis’)
  • using vertex-by-vertex inverse covariance matrix derived from residual timeseries
• “unwhite”: not spatially pre-whitened

Plotting and statistical details:

• within ‘curves’ sections, ribbons indicate 95% CI of bootstrapped mean beta across subjects (one ribbon per run).
• all p-value corrections were performed whole-cortex, with FDR.
• for “stats and maps” sections, HLM was used to estimate t-statistics on the hi-lo contrast per ROI:
  • random-intercept models
  • level-I: (TR within target window)*(run)
  • level-2: subject
  • \(\beta_{trs} = b_0 + x_{tr}b_1 + d_s + \epsilon_{trs}\)
    • \(t, r, s\) are indices for TR, run, and subject
    • \(x\) is dummy-coded predictor for hi condition; \(b_1\) reflects hi-lo contrast (fixed effect)
    • \(d\) is is subject-specific intercept
    • \(\epsilon\) is error

2 Network-level

Estimates are first averaged across parcels within network per subject. Group-level stats and plots are then displayed at the network-level.

2.1 curves

2.1.1 beta contrasts

2.1.2 t-stats on contrast

3 parcel-level

3.1 curves

3.1.1 counts of significant effects per TR

3.1.2 DMCC55B ROIs

3.1.3 largest 20 ROIs

3.2 maps (target TRs)

## $Axcpt
## [1] 260
## 
## $Cuedts
## [1] 0
## 
## $Stern
## [1] 0
## 
## $Stroop
## [1] 490

Unthresholded and thresholded t-statistics displayed on inflated brains. Threshold at fdr-corrected p-value (across all parcels) of 0.05. Note: colors are reverse-scaled, i.e., dark == strong effect, light == weak effect (so strong effects show up on white background).

3.2.1 Axcpt

3.2.1.1 unstand white

3.2.1.2 stand white

3.2.1.3 unstand unwhite

3.2.1.4 stand unwhite

### Cuedts

3.2.1.5 unstand white

3.2.1.6 stand white

3.2.1.7 unstand unwhite

3.2.1.8 stand unwhite

### Stern

3.2.1.9 unstand white

3.2.1.10 stand white

3.2.1.11 unstand unwhite

3.2.1.12 stand unwhite

### Stroop

3.2.1.13 unstand white

3.2.1.14 stand white

3.2.1.15 unstand unwhite

3.2.1.16 stand unwhite